Hi;
Ian, you are correct that I view job submission to a cluster as being
one of the simplest, and hence most basic, HPC use cases to start with.
Or, to be slightly more general, I view job submission to a “black
box” that can run jobs – be it a cluster or an SMP or an SGI NUMA
machine or what-have-you – as being the simplest and hence most basic HPC
use case to start with. The key distinction for me is that the internals
of the “box” are for the most part not visible to the client, at
least as far as submitting and running compute jobs is concerned. There
may well be a separate interface for dealing with things like system
management, but I want to explicitly separate those things out in order to
allow for use of “boxes” that might be managed by proprietary means
or by means obeying standards that a particular job submission client is
unfamiliar with.
I think the use case that Ravi Subramaniam posted to this mailing list
back on 2/17 is a good one to start a discussion around. However,
I’d like to present it from a different point-of-view than he did.
The manner in which the use case is currently presented emphasizes all the
capabilities and services needed to handle the fully general case of submitting
a batch job to a computing utility/service. That’s a great way of
producing a taxonomy against which any given system or design can be compared
to see what it has to offer. I would argue that the next step is to ask
what’s the simplest subset that represents a useful system/design and how
should one categorize the various capabilities and services he has identified
so as to arrive at meaningful components that can be selectively used to obtain
progressively more capable systems.
Another useful exercise to do is to examine existing job scheduling
systems in order to understand what they provide. Since in the real world
we will have to deal with the legacy of existing systems it will be important
to understand how they relate to the use cases we explore. In the same
vein, it will be important to take into account and understand other existing
infrastructures that people use that are related to HPC use cases.
I’m thinking of things like security infrastructures, directory services,
and so forth. From the point-of-view of managing complexity and reducing
total-cost-of-ownership, it will be important to understand the extent to which
existing infrastructure and services can be reused rather than reinvented.
To kick off a discussion around the topic of a minimalist HPC use case,
I present a straw man description of such below and then present a first
attempt at categorizing various areas of extension. The categorization of
extension areas is not meant to be complete or even all that carefully
thought-out as far as componentization boundaries are concerned; it is merely
meant to be a first contribution to get the discussion going.
A basic HPC use case: Compute cluster
embedded within an organization.
·
This is your basic batch job
scheduling scenario. Only a very basic state transition diagram is
visible to the client, with the following states for a job: queued, running,
finished. Additional states -- and associated state transition request
operations and functionality -- are not supported. Examples of additional
states and associated functionality include suspension of jobs and migration of
jobs.
·
Only "standard"
resources can be described, for example: number of cpus/nodes needed, memory
requirements, disk requirements, etc. (think resources that are
describable by JSDL).
·
Once a job has been submitted it
can be cancelled, but its resource requests can't be modified.
·
A distributed file system is
accessible from client desktop machines and client file servers, as well as
compute nodes of the compute cluster. This implies that no data staging
is required, that programs can be (for the most part) executed from existing
file system locations, and that no program "provisioning" is required
(since you can execute them from wherever they are already installed).
Thus in this use case all data transfer and program installation operations are
the responsibility of the user.
·
Users already have accounts within
the existing security infrastructure (e.g. Kerberos). They would like to
use these and not have to create/manage additional authentication/authorization
credentials (at least at the level that is visible to them).
·
The job scheduling service resides
at a well-known network name and it is aware of the compute cluster and its
resources by "private" means (e.g. it runs on the head node of the
cluster and employs private means to monitor and control the resources of the
cluster). This implies that there is no need for any sort of directory
services for finding the compute cluster or the resources it represents other
than basic DNS.
·
Compute cluster system management
is opaque to users and is the concern of the compute cluster's owners.
This implies that system management is not part of the compute cluster's public
job scheduling interface. This also implies that there is no need for a
logging interface to the service. I assume that application-level logging
can be done by means of libraries that write to client files; i.e. that there
is no need for any sort of special system support for logging.
·
A simple polling-based interface
is the simplest form of interface to something like a job scheduling
service. However, a simple call-back notification interface is a very
useful addition that potentially provides substantial performance benefits
since it can enable the avoidance of lots of unnecessary network traffic.
Only job state changes result in notification messages.
·
There are no notions of fault
tolerance. Jobs that fail must be resubmitted by the client.
Neither the cluster head node nor its compute nodes are fault tolerant. I
do expect the client software to return an indication of
failure-due-system-fault when appropriate. (Note that this may also occur
when things like network partitions occur.)
·
One does need some notion of how
to deal with orphaned resources and jobs. The notion of job lifetime and
post-expiration garbage collection is a natural approach here.
·
The scheduling service provides a
fixed set of scheduling policies, with only a few basic choices (or maybe even
just one), such as FIFO or round-robin. There is no notion, in general,
of SLAs (which are a form of scheduling policy).
·
Enough information must be
returned to the client when a job finishes to enable basic accounting
functionality. This means things like total wall-clock time the job ran
and a summary of resources used. There is not a need for the interface to
support any sort of grouping of accounting information. That is, jobs do
not need to be associated with projects, groups, or other accounting entities
and the job scheduling service is not responsible for tracking accounting
information across such entities. As long as basic resource utilization
information is returnable for each job, accounting can be done externally to
the job scheduling service. I do assume that jobs can be uniquely
identified by some means and can be uniquely associated with some principal
entity existing in the overall system, such as a user name.
·
Just as there is no notion of
requiring the job scheduling service to track any but the most basic job-level
accounting information, there is no notion of the service enforcing quotas on
jobs.
·
Although it is generally useful to
separate the notions of resource reservation from resource usage (e.g. to
enable interactive and debugging use of resources), it is not a necessity for
the most basic of job scheduling services.
·
There is no notion of tying
multiple jobs together, either to support things like dependency graphs or to
support things like workflows. Such capabilities must be implemented by
clients of the job scheduling service.
Interesting extension areas:
·
Additional scheduling policies
o
Weighted fair-share, …
o
Multiple queues
o
SLAs
o
...
·
Extended resource descriptions
o
Additional resource types, such as
GPUs
o
Additional types of compute
resources, such as desktop computers
o
Condor-style class ads
·
Extended job descriptions (as
returned to requesting clients and sys admins)
·
Additional classes of security
credentials
·
Reservations separated from
execution
o
Enabling interactive and debugging
jobs
o
Support for multiple competing
schedulers (incl. desktop cycle stealing and market-based approaches to scheduling
compute resources)
·
Ability to modify jobs during
their existence
·
Fault tolerance
o
Automatic rescheduling of jobs
that failed due to system faults
o
Highly available resources:
This is partly a policy statement by a scheduling service about its characteristics
and partly the ability to rebind clients to migrated service endpoints
·
Extended state transition diagrams
and associated functionalities
o
Job suspension
o
Job migration
o
…
·
Accounting & quotas
·
Operating on arrays of jobs
·
Meta-schedulers, multiple schedulers,
and ecologies and hierarchies of multiple schedulers
o
Meta-schedulers
·
Hierarchical job scheduling with a
meta-scheduler as the only entry point; forwarding jobs to the meta-scheduler
from other subsidiary schedulers
o
Condor-style matchmaking
·
Directory services
o
Using existing directory services
o
Abstract directory service
interface(s)
·
Data transfer topics
o
Application data staging
·
Naming
·
Efficiency
·
Convenience
·
Cleanup
o
Program staging/provisioning
·
Description
·
Installation
·
Cleanup
Marvin.
From: Ian Foster
[mailto:foster@mcs.anl.gov]
Sent: Monday, February 20, 2006
9:20 AM
To:
Cc:
Subject: Re: [ogsa-wg] Paper
proposing "evolutionary vertical design efforts"
Dear All:
The most important thing to understand at this point (IMHO) is the scope of
this "HPC use case," as this will determine just how minimal we can
be.
I get the impression that the principal goal may be "job submission to a
cluster." Is that correct? How do we start to circumscribe the scope more
explicitly?
Ian.
At 05:45 AM 2/16/2006 -0800,
Enclosed is a paper that advocates an additional set of activities that
the authors believe that the OGSA working groups should engage in.
Broadly speaking, the OGSA and related working groups are already doing a bunch
of important things:
·
There is broad exploration of the big picture, including
enumeration of use cases, taxonomy of areas, identification of research issues,
etc.
·
There is work going on in each of the horizontal areas that have
been identified, such as
·
There is working going around individual specifications, such as
BES, JSDL, etc.
Given that individual specifications are beginning to come to fruition, the
authors believe it is time to also start defining vertical profilesthat
precisely describe how groups of individual specifications should be employed
to implement specific use cases in an interoperable manner. The authors
also believe that the process of defining these profiles offers an opportunity
to close the design loopby relating the various on-going protocol and standards
efforts back to the use cases in a very concrete manner. This provides an
end-to-end setting in which to identify holes and issues that might require
additional protocols and/or (incremental) changes to existing protocols.
The paper introduces both the general notion of doing focused vertical design
effortsand then focuses on a specific vertical design effort, namely a minimal
HPC design.
The paper derives a specific HPC design in a first principlesmanner since the
authors believe that this increases the chances of identifying issues. As
a consequence, existing specifications and the activities of existing working
groups are not mentioned and this paper is not an attempt to actually define a
specifications profile. Also, the absence of references to existing work
is not meant to imply that such work is in any way irrelevant or
inappropriate. The paper should be viewed as a first abstract attempt to
propose a new kind of activity within OGSA. The expectation is that
future open discussions and publications will explore the concrete details of
such a proposal.
This paper was recently sent to a few key individuals in order to get feedback
from them before submitting it to the wider GGF community. Unfortunately
that process took longer than intended and some members of the community may
have already seen a copy of the paper without knowing the context within it was
written. This email should hopefully dispel any misconceptions that may
have occurred.
For those people who will be around on for the F2F meetings on Friday,
Math & Computer Science Div. Dept of
Computer Science
Argonne National Laboratory The
Tel: 630 252
4619
Fax: 630 252 1997
Globus
Alliance, www.globus.org